So in terms of 'being competitive' on this dataset...

[mdda]

Please let me know whether I've understood the paper correctly:

• Given a set of First-Order-Logic Premises, it's a ~mechanical task to resolve the True/False/Unknown for the given Conclusions
• So, the task that arises naturally from this dataset is a "natural language to first-order logic translation (NL-FOL translation)" task
  • Because, if we can do the NL-FOL translation well enough, then we can rely on a FOL solver to do the 'thinking' for us;

Naively, this seems like a simpler goal than build an NLP model that goes directly Natural Language input to Natural Language resolutions. But I'm guessing that doing the NL-FOL translation task is challenging to do at high enough accuracy rates to enable a solver to work on the outputs?

( Just trying to make sure I don't dive into building a translation tool, whereas the 'real problem' is to do with few-shot prompting, or more super-LLM training-based ).

[ke-lara]

// doing the NL-FOL translation task is challenging to do at high enough accuracy rates to enable a solver to work on the outputs?// Exactly. Semantic parsing is quite challenging to do. Also, taking this line doesn't give any new insights, does it?